CN116244501B - Cold start recommendation method based on first-order element learning and multi-supervisor association network - Google Patents

Abstract

The invention provides a cold start recommendation method based on first-order element learning and a multi-supervisor associated network, which comprises the following steps: s1, converting original features into dense vectors through an embedding layer to obtain dense representations of users and dense representations of objects; s2, inputting dense representation of the user into a multi-supervisor network; s3, inputting the user representation and the dense representation of the object generated by the multi-supervisor network into the association network to calculate the importance of each user and object, and carrying out fine-granularity user-object interaction modeling to obtain the weighted user representation and object representation; s4, inputting the user representation and the item representation generated through fine granularity modeling into a personalized user preference estimation model based on FOMAML, and outputting a final result. The method effectively relieves the problems of excessive model training expenditure and user preference deviation, and improves the accuracy of cold start recommendation of the model.

Description

Cold start recommendation method based on first-order element learning and multi-supervisor association network

Technical Field

The invention relates to the field of cold start recommendation, in particular to a cold start recommendation method based on first-order element learning and a multi-supervisor associated network.

Background

With the rapid development of information technology, the number of internet information is exponentially increased, and the proportion of low-entropy information on the internet is larger and larger, so that the difficulty of obtaining corresponding information according to own needs of internet users is larger and larger. The recommendation system can help users find out information meeting the demands among the complicated information as an effective tool for solving the information overload phenomenon. Conventional recommendation systems can be generally classified into collaborative filtering-based recommendation systems that evaluate user responses by collecting historical preference information of users from a large number of users, and content-based recommendation systems, which are based on ratings of items by other users having similarities to the target user. Such systems cannot handle when there is a new user or new item because of the lack of user-item interaction history and the inability to determine the similarity of the new user or new item to other users or other items. Content-based recommendation systems have been introduced to address the cold start problem, such recommendation systems suggesting items based on the user's profile information and the content of the items, which recommend items to new users and their users with similar content. However, when the user-article interaction history data is sparse, such a system does not perform well in practical use. In addition, because of the improvement of public network security consciousness, the acquisition of personal information of users is more and more difficult, and the recommendation system faces the problem of cold start.

Many efforts are made in the industry to alleviate the cold start problem. In recent years, researchers have introduced optimization-based meta-learning into recommendation systems, and the basic idea of optimization-based parameter initialization is to define meta-knowledge w as initial parameters of a basic recommendation model, and then update parameter initialization in a form of double-layer optimization. In addition to parameter initialization based on a recommendation model, some works also utilize meta-learning to learn adaptive superparameters of different cold start tasks.

The current model based on meta-learning paradigm still has some problems, such as MeLU uses only the relevant attributes of the user and the item, not uses valuable user history interaction sequences, and the embedded vector (embedding) generated by each type of attribute is fixed, which is not considered in the case that the user who may have the same attribute prefers different types of items. MAMO builds a unique embedded vector generator and recommender for each user, but the effect is not very stable, which may be due to overfitting. In addition, the MAML element learning frame is introduced into the multi-element learning model, but partial second derivative operation in the MAML element learning frame can cause the defects of low convergence speed, gradient degradation, high training cost and the like.

In addition, there is a high demand in the industry for time delays in recommendation systems, and in practice the potential preferences of users are elusive, and there may be subtle differences in what they like for users with similar characteristics.

Disclosure of Invention

The invention aims at least solving the technical problems in the prior art, and particularly creatively provides a cold start recommendation method based on first-order element learning and multi-supervisor association network.

In order to achieve the above object of the present invention, the present invention provides a cold start recommendation method based on first-order meta learning and multi-supervisor association network, comprising:

s1, converting original features into dense vectors through an embedding layer to obtain dense representations of users and dense representations of objects; dense representations of users and items are respectivelyWherein->k represents the embedding dimension +.>A dense vector representation representing the nth feature of the user,/->A densified representation of the mth feature of the representation object;

s2, inputting dense representation of the user into a multi-supervisor network;

s3, inputting the user representation and the dense representation of the object generated by the multi-supervisor network into the association network to calculate the importance of each user and object, and carrying out fine-granularity user-object interaction modeling to obtain the weighted user representation and object representation;

S4, inputting the user representation and the item representation generated through fine granularity modeling into a personalized user preference estimation model based on FOMAML, and outputting a final result.

In order to improve the timeliness of the recommendation system and capture the subtle differences of users, the invention introduces a FOMAML element learning framework and constructs a multi-supervisor association network to solve the problems.

Further, a driver gate is employed to guide the supervision forces of each supervisor in the multi-supervisor network.

Further, the driving gate is expressed as:

a ^supervisor ＝softmax(σ(i,Γ(u))) (6)

wherein a is ^supervisor Representing a weight vector of the task item driving gate normalized to the result of the feed-forward neural network using a softmax function; the feedforward neural network is a supervisor network;

sigma () is a three-layer feedforward neural network;

i and u represent item i and user u, respectively; both item i and user u herein are collections.

Γ () is an aggregation function;

final resultIt is the weight vector that may represent the supervising trends of different supervisors.

The user who finally passes through the drive gate is represented as:

r represents a user representation of the dot product operation of the weight vector generated by the driving gate;

a ^supervisor the supervision tendency of different supervisors is represented as a weight vector;

s _i (u) represents a user representation of the ith supervisor after the supervision and instruction of user u;

n represents the number of supervisors.

Each supervisor can supervise different types of articles, the multi-supervisor network can model the article relationship, if the articles are highly correlated, the drive gate will select a similar supervisor network to supervise the user for enhancing the attention, and if the articles are less correlated, the drive gate will learn to supervise the user by using the different supervisor networks.

Further, the aggregation function operates with a mean value. The mean value operation of the user features is to lead in the user information while the model is more focused on the article information.

Further, the step S3 includes:

s3-1, calculating an incidence matrix:

wherein the method comprises the steps ofRepresenting an association matrix;

tanh is a nonlinear activation function;

r represents a user representation of the dot product operation of the weight vector generated by the driving gate;

· ^T representing a transpose of the matrix;

is a weight matrix;

i represents an item i;

incidence matrixRepresenting the degree of association between a corresponding pair of user and item representations at the interaction level.

S3-2, calculating the attention scores of the user and the article by using the characteristics of the association degree of the user-article calculated by the association matrix obtained in the S3-1:

Wherein c _u Representing a user attention score;

c _i representing an item attention score;

W ₁ 、W ₂ representing an attention weight matrix;

by associating momentsArrayThe association network can associate the user and the item to calculate their respective importance.

Normalizing the score using a softmax function yields:

R＝softmax(c _u V ₁ ) (11)

I＝softmax(c _i V ₂ ) (12)

wherein V is ₁ ,V ₂ Representing an importance weight matrix;

r and I represent the estimated user and item importance, respectively.

S3-3, calculating the weighted user representation and the item representation as follows:

wherein the method comprises the steps ofRepresenting the weighted user representation;

representing the weighted representation of the item;

R _n representing an estimated nth user importance;

r _n representing an nth user;

I _m representing an estimated mth item importance;

i _m represents an mth article;

n represents the feature number of the user;

m represents the feature number of the article.

And finally, connecting the weighted user representation and the weighted article representation, and sending the user representation and the article representation into a personalized user preference estimation network for prediction.

Further, the personalized user preference estimation model based on FOMAML comprises a decision layer and an output layer, wherein the decision layer is an N-layer fully connected neural network;

the decision layer is as follows:

the output layer is as follows:

wherein p is ₀ Is the input of the decision layer, and is the overall vector representation of the user representation and the article representation obtained in the step S3;

Representation pair->Performing connection operation;

p ₁ represents p ₀ Vector representations fusing user and item information generated by a first layer of decision layer;

p _N a vector representation representing fused user and item information generated via an N-th decision layer;

· ^T representing a transpose of the matrix;

W _N and b _N Is the weight matrix and the deviation vector of the decision of the N layer;

W _o and b _o Is the weight matrix and the deviation vector of the output layer;

is the user's preference for items;

a and σ represent activation functions of the decision layer and the output layer, respectively. The choice of these activation functions depends on the manner of definition of the user's preferences.

In summary, due to the adoption of the technical scheme, the invention has the following advantages:

(1) The invention provides a cold start recommendation method based on first-order element learning and multi-supervisor association network, which effectively relieves the problem of excessive model training expenditure and user preference deviation and improves the accuracy of cold start recommendation of a model.

(2) In order to alleviate gradient degradation problems caused by slow convergence speed and second-order gradient update of the meta-learning paradigm model on a large data set, the invention introduces a first-order meta-learning framework FOMAML into a recommendation method, and converts second-order derivative operation in the prior meta-learning paradigm into first-order derivative operation.

(3) In order to alleviate the problem of user preference deviation, the invention provides a multi-supervisor associated network learning transferable knowledge for improving user representation quality and extracting a user-object finer granularity representation, wherein a multi-supervisor network is used for generating a user high-quality representation, and then the associated network is used for carrying out fine granularity modeling on the user and the object to enhance the characteristic representation of the user and the object.

(4) The patent of the invention carries out extensive experiments on two reference data sets with different magnitudes, namely MovieLens-100K and bookcross, so as to evaluate the performance of the proposed cold start recommendation method. Experimental results show that the FO-MSAN method is superior to the existing meta-learning paradigm model in cold start recommendation.

Drawings

The foregoing and/or additional aspects and advantages of the invention will become apparent and may be better understood from the following description of embodiments taken in conjunction with the accompanying drawings in which:

FIG. 1 is a schematic diagram of an optimization-based meta learning algorithm.

Fig. 2 is a diagram of the FO-MSAN recommendation framework of the present invention.

Fig. 3 is a graph of the results of a comparison of FO-MSAN and baseline models on ranking tasks.

Figure 4 is a graph of the results of a comparison of FO-MSAN and baseline models in terms of score prediction.

Fig. 5 is a graph of experimental results of the ablation of FO-MSAN on two different network components of a multi-supervisor network and an associated network.

Fig. 6 is a schematic diagram of the impact of different embedded dimensions on FO-MSAN in terms of score prediction.

Fig. 7 is a schematic diagram of the impact of different embedded dimensions on the FO-MSAN approach in terms of ordering tasks.

Detailed Description

Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the drawings are illustrative only and are not to be construed as limiting the invention.

1. Related work

1.1 Meta-learning based on optimization

Meta-learning, also called learning, is to learn, the purpose of which is to train a model that can adapt quickly to a new task with few samples. Meta-learning is mainly inspired by the process of human learning, in which human beings can learn new tasks through fewer learning samples by using knowledge learned before facing the new tasks. Meta-learning can be divided into three different types: metric-based, model-based, and optimization-based meta-learning. In previous meta-learning efforts, metric-based methods learn distance functions on metrics or tasks, while model-based methods aim at designing an architecture or training process to rapidly summarize tasks. Finally, the optimization algorithm is directly adjusted based on the optimization method. In recent years meta-learning frameworks have attracted interest in many areas such as recommendations, natural language processing, and computer vision. In the prior art, metric-based and model-based meta-learning is mainly applied to classification problems, but different personalized commodities are provided for users according to personal needs of the users in a recommendation system, so that optimized meta-learning is introduced into the recommendation system. The meta-learning based on optimization mainly learns various kinds of knowledge in a recommendation system through cross-task, namely global sharing initialization of model parameters, and the model can be more quickly adapted to a new task with a small amount of task data through learning a global initialization parameter, so that a large number of experiments of many researchers have proved that the meta-learning has a good effect in improving the cold start of the recommendation system.

1.FOMAML

The current meta-learning framework is of different types, such as MAML, FOMAML, reptile, etc. MAML can be applied to different models and even different task types in methodology because current machine learning methods perform gradient updates, and MAML's focus is on gradient updates. The core idea of MAML is very simple: if the data is divided into meta_train and meta_test, wherein meta_train contains data of a plurality of tasks, it can be divided into a Support set (Support set) and a Query set (Query set) for training and testing, respectively. In each iteration, there is an initial parameter θ, and gradient updating is performed by using the support set for k tasks, so as to obtain new parameters θ corresponding to different tasks _i ' the global initial parameters θ are then updated over k tasks using the query set, and an optimization-based meta-learning algorithm is shown in fig. 1.

The grey branch lines in the figure represent the update directions of the different tasks, while the black main axes represent the final trend of the model parameters, which prevents the parameters from being overfitted to a certain task. The last dotted line represents the adaptation to the new task and also the fine tuning of the model parameters. The adaptation process may be a simple several step gradient update to adapt the model to the new task. MAML, however, computes the gradient twice during each iteration, which means that more computation time is required.

FOMAML is proposed to solve the problem that the second iteration in MAML requires more computation time, and FOMAML does not perform the second gradient update. In MAML, the optimization objective is:

minimize _θ representing the search for a key L _τ (. Cndot.) the initialization parameter θ with minimum loss;

representing a desire on task τ;

L _τ (-) represents the corresponding loss function on task τ;

wherein the method comprises the steps ofThe parameter update, that is, the gradient update is performed k times on k samples of the global initial parameter θ using the task τ, so the gradient is:

representing update operation->Jacobian matrix of (a);

g _MAML representing the gradient of MAML;

L′ _τ (-) represents the derivative of the loss function corresponding at task τ;

θ _new representing the parameters obtained after gradient updating, corresponding to the addition of a sequence of gradient vectors, i.e. θ, to the initial parameter vector θ _new ＝θ+g ₁ +g ₂ +...+g _k ，g _i Representing the gradient update results of different tasks.

Wherein the method comprises the steps ofThe second derivative is involved and can be relatively time consuming. FOMAML is no longer divided into k different θ's when gradient calculations are performed on k tasks' _i ' i indicates what number of tasks is, but the gradient calculated by the previous task is used in the gradient calculation of the next task, and the final global gradient update is also the parameter theta obtained after the k tasks are calculated _i Updating, wherein the gradient is as follows:

g _FOMAML represents the gradient of FOMAML;

L′ _τ (-) represents the derivative of the loss function corresponding at task τ; θ _i Representing the parameters obtained after gradient update, and θ _new Differently, θ _i Corresponding to the addition of a constant gradient vector sequence to the initial parameter vector θ.

Experimental results show that the final effect of MAML and FOMAML is almost the same, but gradient update is changed from second-order operation to first-order operation, so that the calculation cost is saved. Therefore, the invention introduces the FOMAML element learning framework as a parameter updating mode of the cold start recommending method FO-MSAN, and effectively reduces the cost during model training.

1.3 Cold Start recommendation

The recommendation algorithm based on deep learning shows superiority in improving recommendation precision. However, the deep learning-based recommendation algorithm makes it difficult to recommend a new user and a new item to be presented after the predictive model training due to the limited historical data and the deviation of user preference caused by personal differences. The above cold start problem is generally present in recommended systems. To alleviate the lack of historical data, auxiliary information is used in the recommendation system to describe new users and new items. For example, bansal et al apply word-level embedding to represent new items of semantic features extracted from their descriptive sentences. Li et al describe a behavior-intensive neural network that also utilizes the textual content of items to learn semantic potential item vectors and represent new users by aggregating their interacted with items. In addition to the text feature-based recommendation system described above, cheng et al utilized sparse features and proposed Wide & Deep to learn the linear and Deep relationships of these features together. To better learn the relationship between feature embedding, cheng et al developed an adaptive decomposition network (AFN). AFN proposes a logarithmic transformation network to learn arbitrary order cross-over features from data efficiently. However, even though users have different historical interaction data, the content-based recommendation system described above always provides the same suggestions for users having the same side information. These recommendation systems ignore individual differences in user interactions and fail to accurately provide personal recommendations. At the same time, cross-domain recommendations (CDRs) utilize data from external domains to provide insufficient interactions in the target domain. CDRs are seen as a promising solution to alleviate the lack of interaction problem. However, the CDR is used to require the data set to exist in the external domain shared by users or shared by context functions, most of the data sets do not exist in the relationship, in order to make the cold start recommendation model more universal, the invention adopts the concept of meta learning, and replaces the common MAML meta learning framework, so that the FOMAML meta learning framework is introduced into the model of the invention, and the cost of the model in training is reduced.

2. The cold start recommendation method

2.1 Total frame

In the present patent, the problem to be solved by the present method can be defined as follows: assuming that there are three sets U, I, R, U represents a user (item) set, I represents an item set, and R represents a user-item interaction set. For each user U e U, it is necessary to predict for each itemIs +.>Wherein->Representing a set of items that have not interacted with by user u. This preference score->By user-item model->The calculation results are that:

and finally, sorting according to the preference scores, and recommending the first K items in the sorted list to the user u. Suppose a userx represents the feature, N represents the feature number of the user, < ->Representing the nth characteristic of the user, the articleWherein M represents the characteristic number of the article, +.>Representing the mth characteristic of the article. Using the user and item content characteristics as inputs, a cold start recommendation framework is designed using a cold start recommendation method based on first order element learning and multi-supervisor association network, the final FO-MSAN recommendation framework formed is shown in fig. 2.

In fig. 2, the overall steps performed by the FO-MSAN recommendation framework are as follows: first the original features are transformed by the embedding layer into a dense vector e, the representation of the user and the item becomes Wherein->k represents the embedding dimension +.>A dense vector representation representing the nth feature of the user,/->A densified representation of the mth feature of the representation object. The dense representation of the user is then input into a multi-supervisor network where the single supervisor is limited in attention and the field of view covers only a portion of the user's information and does not fully cover the user's features. Considering that the given degree of supervision by the supervisor should be different, a driver gate is set to instruct the supervisor how the degree of supervision should be given. The user representations generated by the multi-supervisor network and the dense representations of items are then input to the association network to calculate the importance of each user and item, and a fine-grained user-item interaction modeling is performed. Finally, the user and item representations generated through fine-grained modeling are input into a personalized user preference estimation model based on FOMAML to output final results. In terms of parameter updating, the FO-MSAN recommendation framework will update the decision layer and output layer (marked with blue boxes) according to the support and update of each user, and after local updates for the user, all modules (marked with purple money) will update globally according to the query and.

2.2 Multi-supervisor network

To enhance the recommendation capabilities of a recommendation system, high quality user representations must be generated to represent user intent for different items. A supervisor network (a feed-forward network) may typically be employed to supervise the tendencies of the generated user representations on the items. But the field of view of the individual supervisors is limited and does not provide adequate supervision and instruction. Thus, a multi-supervisor network is employed to supervise and guide the user. Formally, all supervisor-directed user representations are averaged:

wherein r represents an average of the user representations after the plurality of supervisor guides;

s _i (u) represents a user representation of the ith supervisor after the supervision and instruction of user u;

s _i representing an ith supervisor;

n represents the number of supervisors.

But such averaging may affect certain information about an item to some extent, resulting in the resulting user representation r being insensitive to certain item tasks. The degree to which different supervisors actually understand the task of the item will vary, as will the degree to which they are adept. It is now desirable for a supervisor to be able to supervise the tendencies expressed by the user over his task of good, rather than to generalize globally, all without being precise.

In accordance with this need, the present patent contemplates a driver gate that directs the supervisory process with item characteristics as input, but the supervisor's understanding of the different item tasks varies, as does the processing power of the generated user representation. Thus, the driver gate should take as input, in addition to the item characteristics, also user characteristics to ensure that the supervisor network can properly supervise the generated user representations' tendencies to the item. Under this definition, a task item driven gate may be represented as:

a ^supervisor ＝softmax(σ(i,Γ(u))) (6)

wherein a is ^supervisor The weight vector representing the normalization of the task item driver gate to the results of the feed forward neural network using the softmax function. The single supervisor network is a multi-layer feed forward neural network, and the multi-supervisor network is a collection of multiple single supervisor networks.

Sigma () is a simple feed-forward neural network;

i and u are feature inserts for item i and user u, which is a dense representation; the softmax function is used to normalize the sigma () output.

Γ () is an aggregation function, where mean operations are employed, which mean user features to introduce user information while wanting to make the model focus more on item information.

Final resultIt is the weight vector that may represent the supervising trends of different supervisors.

The user representation at the last pass through the drive gate is restated as:

r represents a user representation of the dot product operation of the weight vector generated by the driving gate;

a ^supervisor the supervision tendency of different supervisors is represented as a weight vector;

s _i (u) represents a user representation of the ith supervisor after the supervision and instruction of user u;

n represents the number of supervisors.

Each supervisor can supervise different types of articles, the multi-supervisor network can model the article relationship, if the articles are highly correlated, the drive gate will select a similar supervisor network to supervise the user for enhancing the attention, and if the articles are less correlated, the drive gate will learn to supervise the user by using the different supervisor networks.

2.3 associated networks

The primary purpose of the association network is to build a more fine-grained model representation of the item and user representations. The present patent designs an association network to computer the importance of each user and item. In particular, the associated network uses the impact of the user representation after passing through the multi-supervisor network to estimate the importance of the item, and as such, uses the impact of the item to estimate the importance of the user. In order to realize that the object representation is introduced when the importance of the user is calculated or the user representation is introduced when the importance of the object is calculated, an association matrix is adopted in the calculation process. The user representation that has been calculated before may be The article is denoted +.>Then this association matrix->The calculation mode of (2) is as follows:

wherein the method comprises the steps ofIs a weight matrix;

tanh is a nonlinear activation function;

r represents the final generated r through a plurality of networks, namely, the user representation of the dot product operation of the weight vector generated through the driving gate;

· ^T representing a transpose of the matrix;

incidence matrixThe degree of association between a corresponding pair of user and item representations at the interaction level is represented.

After this correlation matrix is calculated, it can be used as a feature of neural network to calculate the degree of user-item correlation to calculate the user/item attention score:

wherein c _u Representing a user attention score;

c _i representing an item attention score;

W ₁ 、W ₂ representing an attention weight matrix;

by an association matrixThe model may correlate the user and item to calculate their respective importance and the final weight of the final user-item importance may be normalized to the score using the softmax function:

R＝softmax(c _u V ₁ ) (11)

I＝softmax(c _i V ₂ ) (12)

wherein V is ₁ ,V ₂ Representing an importance weight matrix;

r and I represent the estimated user and item importance, respectively.

The weighted user representation and item representation can thus be calculated as:

R _n representing an estimated nth user importance;

r _n Representing an nth user;

I _m representing an estimated mth item importance;

i _m represents an mth article;

n represents the feature number of the user;

m represents the feature number of the article;

and finally, connecting the processed user representation and the processed article representation, and sending the user representation and the article representation into a personalized user preference estimation network for prediction.

2.4 personalized user preference estimation model based on FOMAML

The personalized user preference estimation model is composed of a decision layer and an output layer, wherein the decision layer is essentially an N-layer fully-connected neural network, the output layer is a subsequent layer of the decision layer, and the output layer outputs a score to describe the probability of user and article interaction. These layers may be configured to:

wherein p is ₀ Is an overall vector representation that links the user representation and the item representation;

representation pair->Performing connection operation;

· ^T representing a transpose of the matrix;

W _N and b _N Is the weight matrix and the deviation vector of the decision of the N layer;

W _o and b _o Is the weight matrix and the deviation vector of the output layer;

is the user's preference for items;

a and σ represent activation functions of the decision layer and the output layer, respectively. The choice of these activation functions depends on the manner of definition of the user's preferences.

Inspired by first-order element learning, the invention introduces FOMAML into a personalized user preference estimation network, and a parameter gradient updating mode based on the FOMAML is shown as an algorithm 1. The gradient updating method adopts SGD or Adam.

/>

Fig. 2 is a specific process of using FOMAML for parameter updating in combination with algorithm 1 by the FO-MSAN recommendation framework, using the user's item consumption history and preferences as a support set for local updating during model local updating, while not updating the vector representations of the user and item to ensure the stability of the learning process.

4. Experimental analysis

A number of comparative analytical experiments will be performed next, with the aim of answering the following research questions:

problem 1: is the proposed cold start recommended FO-MSAN method superior to other algorithms in practical testing?

Problem 2: how does the individual network modules in the FO-MSAN affect the overall model?

Problem 3: how does the superparameter set up to affect the FO-MSAN?

4.1 data sets

The invention adopts two reference data sets of MovieLens-100K and Bookcrossing to verify the FO-MSAN algorithm. Both data sets provide basic user and item information, such as the age of the user and the date of publication of the item, while both data sets have explicit feedback information, and table 1 summarizes the characteristics of both data sets.

Table 1 basic statistics and usage of MovieLens-100K and Book cross data sets

In addition, the large difference between the data volumes between the two data sets can also be used as a contrast between the small data set and the large data set. The processing of the two data sets is as follows:

MovieLens-100K MovieLens data set contains rating data of a plurality of users for a plurality of movies, metadata information of movies and user attribute information are also included, movieLens-100K is a small data set of 100000 interactions which is constructed by randomly extracting from a well-known reference data set of MovieLens, the data set is formatted into tasks in preprocessing, and each user is individually represented as a task. Simultaneously dividing training sets: verification set: the task ratio of the test set is 8:1:1. For each task, 10 interactions are randomly selected as a Query set (Query set) and the other interactions as Support sets (Support set). To more fully verify the performance of the model, performance tests were performed on both scoring prediction and ranking tasks for the model using Movie-100K, respectively. For the scoring settings, raw scoring data is used. And setting positive labels with scores greater than or equal to 4 and negative labels with scores less than 4 for ranking task setting.

Bookcrossing is a book scoring dataset, which is formatted as tasks, each user being represented as a task separately, as in the MovieLens-100K dataset preprocessing operation. While each dataset was divided into 8:1:1 ratios for training, testing and validation, respectively. For each task, 10 interactions are randomly selected as a query set (query set) and the other interactions as support sets (support sets). As with the MovieLens-100K dataset, the model was tested for performance in both scoring prediction and ranking tasks, respectively, using Bookcrossing. For the scoring settings, raw scoring data is used. For ranking task setting, positive labels are set for scores greater than or equal to 8, and negative labels are set for scores less than 8.

4.2 evaluation index

The performance of the FO-MSAN method is verified from two aspects of scoring prediction and sequencing tasks, and the performance evaluation index of scoring prediction adopts average absolute error (MAE):

wherein the method comprises the steps ofRepresenting the predicted score value, y _ij The real scoring value is represented, N is the number of samples, and the smaller the MAE evaluation index is, the better the model effect is.

To evaluate the ranking quality of the items, the performance of the model was quantified using three performance metrics, normalized break cumulative gain (Normalized Discounted Cumulative Gain), average reciprocal ranking (Mean Reciprocal Rank), and accuracy (Precision).

The expression for ndcg@n is as follows:

dcg@n is a break cumulative gain, and its expression is:

where rel (i) represents the degree of relatedness of an item at position i.

IDCG is an ideal DCG, and its expression is:

where |REL| denotes a set of top N results ordered by relevance size.

MRR@N refers to the reciprocal of the ranking of a plurality of items at the correct position, and the formula is as follows:

wherein rank is _i Representing the rank of the first correct position of the ith item, N representing the top N results.

precision@N indicates how many proportions in the final recommendation list are user item interaction records that have occurred:

Wherein L is _N Is the set of N items actually recommended to the user, T represents the top N items in the model predictive recommendation list.

4.3 comparison method

In order to evaluate the rationality and effectiveness of the proposed FO-MSAN method, a comparison analysis was performed with 8 classical algorithms, neuMF, NNCF, etc.

Deep learning based model:

(1) NeuMF. NeuMF proposes a neural network to model potential features of users and projects, and utilizes multi-layer perceptrons to impart a high level of non-linear modeling capability to the model.

(2) NNCF. NNCF integrates neighborhood information into neural collaborative filtering method, and in this way, user-commodity interaction data is supplemented, so that the effectiveness of the model on recommended tasks is enhanced.

(3) The NGCF establishes a high-order connectivity embedding layer, refines the embedding of the user by aggregating the embedding of the interaction items, and the superposition multiple embedding propagation layers capture the high-order connectivity of the collaboration signals and integrate the high-order connectivity into the predictive model.

(4) Lightgcn.lightgcn counteracts both the feature transformation and the nonlinear activation in the GCN, learns user and item embeddings by linear propagation on the user and item interaction graph, and uses the weighted sum of the embeddings learned at all layers as the final embeddings, making the model easier to implement and train.

Model based on meta learning:

(1) The melu learns a global initialization parameter through the neural network as a priori knowledge that is used to generate a single model for each user to provide user-specific suggestions.

(2) LWA.LWA implements a linear classifier based on meta-learning strategy, with its weights determined by project history, and its test results are significant due to MF.

(3) Tanp.tanp is able to learn the relevance of different tasks and customize the global knowledge learned by meta-learning to parameters related to the tasks to estimate user preferences, which treats each task as a random process. From historical behaviours of the user to conditional priors, the historical behaviour is mapped to a predictive distribution.

(4) MAMO the method improves on MeLU, and adds User Profile-aware memory network to store User, article and task information, and uses the information to provide personalized initialization parameters for User.

4.4 Experimental Environment and parameter settings

In the experimental link, the experiment is configured as Windows11, CPU i9-10980XE, 64GB memory and NVIDIA Quadro RTX 5000, and the environment required by the experiment is Python is more than or equal to 3.7.0, recbole is more than or equal to 1.1.1, numpy is more than or equal to 1.20.3, torch is more than or equal to 1.11.0, and tqdm is more than or equal to 4.62.3.

The model of patent comparison of the invention comprises a model based on deep learning and a model based on meta-learning paradigm on experimental parameter setting. The two models are provided with larger difference in learning rate, the model based on deep learning has only one learning rate, and the model based on meta learning has two learning rates, namely a global learning rate and a local learning rate. For the model based on deep learning, the learning rate was uniformly set to 0.01. For the model based on meta learning, the local learning rate is set to 5×10 ^-6 The global learning rate is set to 5×10 ^-5 。

4.5 experimental results and analysis

4.5.1 comparative experiments

In order to verify the improvement effect of the proposed cold start recommendation method FO-MSAN compared with a baseline model on a sequencing task, a comparison experiment of the FO-MSAN and the baseline model is carried out, N=5 is taken as indexes MRR@N, NDCG@N and precision@N, and the average value of 5 experimental results is taken as a final experimental result. Using the above criteria, the MRR, NDCG, precision results for the different models under the different data sets are shown in table 2.

TABLE 2 Experimental results of sequencing tasks on MovieLens-100K and Bookcrossing datasets, bolded are the best index results, italicized are the next best index results

As can be seen from table 2, on the MovieLens-100K dataset, the FO-MSAN proposed by the present invention is superior to the comparative model in each index, wherein the mrr@5 index is 0.7618, the result is improved by 3% compared to the MeLU model, the ndcg@5 index is 0.6371, the result is improved by 4.42% compared to the MAMO model, the precision@5 index is 0.6269, and the result is improved by 1.63% compared to the MAMO model. On a Bookcrossing dataset, FO-MSAN is superior to a comparison model in NDCG@5 and precision@5 indexes, wherein the MMR@5 index is only slightly lower than the MAMO model, the result of the NDCG@5 index is improved by 1.13% compared with the best result in the comparison model, and the result of the precision@5 index is improved by 1.07% compared with the best result in the comparison model. The overall comparison of performance indexes of the ranking tasks is shown in fig. 3, and the result shows that the association network can effectively improve the accuracy of the ranking tasks by utilizing the user-article information with finer granularity to enhance the characteristic representation strength, and particularly, the effect on a small data set is better, the multi-supervisor network enables a model to extract various article specific information, and the model is better helped to generate high-quality user representation.

The basic information of the data set can be derived from table 1, and it can be seen that MovieLens-100K has a great difference between the data amount and the bookcross, so that the improvement effect of FO-MSAN on a large data set tends to be stable, and an abrupt improvement occurs on the improvement effect of a data set with a smaller data amount, but the effect of the method proposed in the patent of the invention is improved compared with the baseline model regardless of the size of the data set. Meanwhile, the experimental results show that the results of the method based on the deep learning on two data sets are different from those of the method based on the meta learning, which is related to the setting of evaluation indexes, the table 2 is the experimental results of the sorting tasks, the 3 evaluation indexes are all related to the sorting tasks, and the method based on the deep learning in the baseline model is biased to the scoring prediction. To verify the performance of the FO-MSAN method in terms of score prediction, the present patent conducted experiments using MAE index, and the experimental results are shown in table 3.

TABLE 3 Experimental results of scoring predictions on MovieLens-100K and Bookcrossing datasets, best results for the deep learning model are underlined, best results for the meta-learning model are bolded

As can be seen from the data in table 3 and fig. 4, FO-MSAN is superior to the meta-learning based approach in terms of score prediction, 2.38% improvement over 0.7661 of MeLU on MovieLens-100K dataset and 2.38% improvement over 0.7661 of MeLU on bookcross dataset. This benefits from the excellent feature extraction and learning capabilities of the multi-supervisory network, enabling the FO-MASN to generate high quality user representations to promote model recommendation. However, compared with the deep learning-based method, the method of the invention has the advantages that the effect of FO-MSAN is not the best except for the two methods of NGCF and LightGCN based on the deep learning, which shows that the method structure of the invention is more prone to ordering tasks, but has the function of the spam of a multi-supervisor network, and the FO-MSAN has great improvement on the scoring and predicting task compared with the model based on the deep learning and the meta-learning paradigm model.

4.5.2 ablation experiments

In order to better understand the contributions of the different network structures in the FO-MSAN, the present patent conducted ablation experiments on both data sets. Two variants of FO-MSAN were formulated: (1) A model without multi-supervisor network, named FO-MSAN-MS; (2) a model without associated networks, named FO-MSAN-AN. The super parameter settings of the two are the same, the dimension setting 32 is embedded, and the local update learning rate is 5 multiplied by 10 ^-6 Global update learning rate of 5×10 ^-5 . In order to comprehensively consider the influence effects of different network structures on both scoring prediction and sequencing tasks, the MAE index is adopted to evaluate the scoring prediction aspect, and the NDCG@5 index is adopted to evaluate the sequencing tasksThe results of the comparison of the valence and final ablation experiments are shown in fig. 5.

In fig. 5, it can be observed that two different network structures play a role in gain on both data sets, and that MovieLens-100K is greater than bookcross in gain amplitude, while it can be observed that the influence of the associated network on the model effect on the MovieLens-100K data set is greater than that of the multi-supervisor network. This may be because there is insufficient data for the multi-supervisor network to learn sufficient transferable knowledge for the small data set, while the associated network, because it is insensitive to the amount of data, its feature enhancement can provide more gain effects on the small data set. Comparing with the effects on the bookcross dataset, it can be found that the multi-supervisor network and the associated network provide a gain effect on the whole model that is comparable when the data is sufficient.

4.5.3 parameter sensitivity analysis

Finally, the patent of the invention performs parameter sensitivity analysis on both scoring prediction and sequencing tasks. It is mainly discussed how the model behaves in different feature embedding dimensions. The values of the embedding dimensions are 8, 16, 32, 64, 128 and 256, the comparison results of the different embedding dimensions in the aspect of grading prediction are shown in fig. 6, and the comparison results of the different embedding dimensions in the aspect of sequencing tasks are shown in fig. 7.

In fig. 6, it can be seen that when the embedding dimension is taken to be 32, the value of MAE is at the lowest point of the whole curve, and the values on the left and right sides of the MAE are far greater than the value of the point, which indicates that selecting the embedding dimension to be 32 on the scoring prediction task can better embody the effect of the FO-MSAN method.

It can be seen in fig. 7 that the FO-MSAN, while not performing best on mrr@5, performs well on ndcg@5 and precision@5 indices when the embedded dimension is taken to 32, whereby it can be determined that the overall performance of the FO-MASN in terms of ordering tasks is best when the embedded dimension is taken to 32, and that the performance of the model is more stable around this dimension, indicating that the FO-MSAN approach is robust.

In summary, the present invention proposes a cold start recommendation method (FO-MSAN) based on first order element learning and multi-supervisor association network. The method introduces the idea of meta learning so that the model can estimate the user preference according to a small amount of articles, and simultaneously considers the complexity of the gradient updating algorithm, introduces the FOMAML framework into the method, thereby reducing the cost of model training. In order to ensure the robustness of the model, the adaptability to various articles and the learning of potential user clusters to eliminate the overfitting condition, a multi-supervisor network is designed to ensure that the generated user representation can have tendencies on different articles, and meanwhile, the association network of the model can give finer-granularity representation on the scale of the user-articles, so that the characteristic representation of the user and the articles is enhanced, and the model recommendation effect is improved. To assess the effectiveness of FO-MSAN, extensive experiments were performed on both the MovieLens-100K and Bookcrossing data sets. Experimental results show that the FO-MSAN model synthesis is superior to the baseline model used in the present patent, especially in terms of ordering tasks. In addition, the different sizes of the data sets used in the ablation experiment also laterally verifies that the characteristic enhancement effect of the associated network module of the FO-MSAN is good on the small data set, but the effect of the characteristic enhancement effect on the small data set is not obvious because the multi-supervisor network does not have enough data to learn transferable knowledge; on a large data set, the comprehensive effect of the multi-supervisor network and the associated network can well improve the effect of the model. The final parameter sensitivity analysis also shows that the effect of the whole model presents a certain stability when the embedding dimension of the model is larger than a certain value, and the proposed FO-MSAN method is robust.

While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.

Claims (4)

1. A cold start recommendation method based on first order meta learning and multi-supervisor associated network, comprising:

s1, converting original features into dense vectors through an embedding layer to obtain dense representations of users and dense representations of objects;

s2, inputting dense representation of the user into a multi-supervisor network;

s3, inputting the user representation and the dense representation of the object generated by the multi-supervisor network into the association network to calculate the importance of each user and object, and carrying out fine-granularity user-object interaction modeling to obtain the weighted user representation and object representation;

s3-1, calculating an incidence matrix:

wherein the method comprises the steps ofRepresenting an association matrix;

tanh is a nonlinear activation function;

r represents a user representation of the dot product operation of the weight vector generated by the driving gate;

representing a transpose of the matrix;

is a weight matrix;

i represents an item i;

s3-2, calculating the attention scores of the user and the article by using the characteristics of the association degree of the user-article calculated by the association matrix obtained in the S3-1:

wherein c _u Representing a user attention score;

c _i representing an item attention score;

W ₁ 、W ₂ representing an attention weight matrix;

normalizing the score using a softmax function yields:

R＝softmax(c _u V ₁ ) (11)

I＝softmax(c _i V ₂ ) (12)

wherein V is ₁ ,V ₂ Representing an importance weight matrix;

r and I represent estimated user and item importance, respectively;

s3-3, calculating the weighted user representation and the item representation as follows:

wherein the method comprises the steps ofRepresenting the weighted user representation;

representing the weighted representation of the item;

R _n representing an estimated nth user importance;

r _n representing an nth user;

I _m representing estimated mth item importance；

i _m Represents an mth article;

n represents the feature number of the user;

m represents the feature number of the article;

s4, inputting the user representation and the item representation generated through fine granularity modeling into a personalized user preference estimation model based on FOMAML to output a final result;

the personalized user preference estimation model based on FOMAML comprises a decision layer and an output layer, wherein the decision layer is an N-layer fully-connected neural network;

the decision layer is as follows:

p ₁ ＝a(W ₁ ^T p ₀ +b ₁ ),

...

The output layer is as follows:

wherein p is ₀ Is the input of the decision layer, and is the overall vector representation of the user representation and the article representation obtained in the step S3;

representation pair->Performing connection operation;

p ₁ represents p ₀ Vector representations fusing user and item information generated by a first layer of decision layer;

p _N representation generation via an N-layer decision layerA vector representation of the converged user and item information;

representing a transpose of the matrix;

W _N and b _N Is the weight matrix and the deviation vector of the decision of the N layer;

W _o and b _o Is the weight matrix and the deviation vector of the output layer;

is the user's preference for items;

a and σ represent activation functions of the decision layer and the output layer, respectively.

2. The method of claim 1, wherein a driver is employed to guide the degree of supervision of each supervisor in the multi-supervisor network.

3. A cold start recommendation method based on first order meta learning and multi-supervisor associated network as claimed in claim 2, wherein said driver gate is expressed as:

a ^supervisor ＝softmax(σ(i,Γ(u))) (6)

wherein a is ^supervisor Representing a weight vector of the task item driving gate normalized to the result of the feed-forward neural network using a softmax function; the feedforward neural network is a supervisor network;

Sigma () is a three-layer feedforward neural network;

i and u represent item i and user u, respectively;

Γ () is an aggregation function;

the user who finally passes through the drive gate is represented as:

r represents a user representation of the dot product operation of the weight vector generated by the driving gate;

a ^supervisor the supervision tendency of different supervisors is represented as a weight vector;

s _i (u) represents a user representation of the ith supervisor after the supervision and instruction of user u;

n represents the number of supervisors.

4. A cold start recommendation method based on first order meta-learning and multi-supervisor associated networks as claimed in claim 3, wherein said aggregation function operates with a mean value.